Set up an assistant with your own instructions

Write it once, not in every chat

If you paste the same background into every chat (what you sell, who your customers are, how you like things written), save it instead. Most of the big assistants, such as ChatGPT, Claude and Gemini, let you store standing instructions and reference files. The feature’s name (custom instructions, projects, custom assistants), where it lives and which plans include it all vary and change, so check your own.

There are usually two levels: preferences that apply to every chat, and a separate assistant or project for one job, with its own instructions and files. Use the first for how you like answers, the second for work.

Make one assistant per job, not one for everything. A plumbing firm’s “booking replies” assistant, with the price list and call-out policy attached, is easy to test and fix. An assistant for everything is hard to judge at anything.

What goes in the instructions

Keep them under a page. Say what to do, not only what to avoid, and be specific enough that a new hire could follow them.

• The job: “You draft replies to booking questions for a plumbing firm.”
• Rules: never promise arrival times, discounts or refunds; quote only prices from the price list.
• Tone: three words, plus one reply you were proud of.
• Sources: answer only from the attached files; if the answer isn’t there, say so and name who to ask.
• Output: length, format, and [CHECK] beside anything uncertain.

Reference files: current, few and clean

Files are how it knows your facts: a price list, a policy page, an answer bank. Keep one current version of each, named clearly, with a “last updated” date inside, and remove old versions. With two price lists attached, it can quote either.

Small, focused files beat one huge manual: with big files, an assistant may only read the parts that look relevant and miss a detail. Keep out customer details and passwords, and check your plan’s data settings for how files and chats are stored and whether they can be used for training.

The three questions, for an assistant

An assistant isn’t an automation, but its instructions often become the AI step in one, so answer the three questions now.

• Who approves: a person reads every draft before it reaches a customer.
• If nobody does: nothing goes out. An assistant in a chat window can’t send anything unless you connect it to your email or other tools, and if you do, keep sending manual.
• How you’ll know it broke: it quotes an old file or fills a gap with a guess. Keep five real questions with known answers, one of them outside the files, and rerun them after every change. If you correct the same thing twice, fix the instructions.